Segmentation-less and Non-holistic Deep-Learning Frameworks for Iris

Recognition

Hugo Proenca and Joao C. Neves

IT - Instituto de Telecomunicacoes

University of Beira Interior, Portugal

{hugomcp,jcneves}@di.ubi.pt

Abstract

Driven by the pioneer iris biometrics approach, the most

relevant recognition methods published over the years are

”phase-based”, and segment/normalize the iris to obtain di-

mensionless representations of the data that attenuate the

differences in scale, translation, rotation and pupillary di-

lation. In this paper we present a recognition method that

dispenses the iris segmentation, noise detection and nor-

malization phases, and is agnostic to the levels of pupillary

dilation, while maintaining state-of-the-art performance.

Based on deep-learning classification models, we ana-

lyze the displacements between biologically corresponding

patches in pairs of iris images, to discriminate between

genuine and impostor comparisons. Such corresponding

patches are firstly learned in the normalized representa-

tions of the irises - the domain where they are optimally

distinguishable - but are remapped into a segmentation-less

polar coordinate system that uniquely requires iris detec-

tion. In recognition time, samples are only converted into

this segmentation-less coordinate system, where matching

is performed. In the experiments, we considered the chal-

lenging open-world setting, and used three well known data

sets (CASIA-4-Lamp, CASIA-4-Thousand and WVU), con-

cluding positively about the effectiveness of the proposed

algorithm, particularly in cases where accurately segment-

ing the iris is a challenge.

1. Introduction

After almost three decades of research [4], iris recogni-

tion under controlled conditions tends to be considered a

solved problem, with most efforts being now concentrated

in augmenting the recognition robustness [2].

Deep-learning frameworks were unquestionably a break-

through in many computer vision tasks, from image seg-

mentation [16], to object detection [32] and classifica-

tion [14]. Being data-driven models, these frameworks do

Scleric Bound. Param. 2

Noise-free Detection 3

Pupillary Bound. Param. 1

Iris Detection. 1

Proposed Method✓State-of-the-Art

Normalized (Noise-Free)

Representation

Segmentation-less

Polar Representation

Normalized Cartesian Polar

Learning Test

[x1,x2] [x′

1,x′

2] [x′′

1,x′′

2]

fn→c(x) fc→p(x)

Figure 1. We propose a non-holistic iris recognition method that

doesn’t require iris segmentation and is agnostic to the levels of

pupillary dilation. During learning time, we infer the appearance

of biologically corresponding patches in segmentation-less repre-

sentations of the iris data, which are the uniques used in recogni-

tion (test) time.

not depend on human efforts to specify image features,

upon the availability of large amounts of data. In the

biometrics context, many deep-learning based recognition

methods have been proposed, based in the face [6] [10],

iris/ocular [39] [26] or finger vein [27] traits. Despite the re-

markable ability of this kind of frameworks to model com-

plex data patterns, demanding them to autonomously infer

the concept of identity might be too ambitious, particularly

when working with poor-quality data.

1

This paper describes one iris recognition algorithm that

- singularly - uses deep learning frameworks exclusively

to perceive what are biologically corresponding patches in

pairs of images, which is an evidently easier task than in-

ferring the identity at once. In this sense, our work can be

regarded as an evolution of the concept behind IRINA [25],

with one major novelty: in test (recognition) time, we don’t

even segment the pupillary and scleric iris boundaries nor

discriminate the occluded parts of the iris. Instead, we rely

uniquely in one iris detection step to obtain a segmentation-

less representation of the iris data, where all subsequent

analysis is carried out.

Note that previous works have also proposed

segmentation-less algorithms, but they were specifi-

cally designed for different traits, such as the ocular region

(e.g., [28] and [29]), where changes in signals’ phase

are not as problematic as in the case of the iris. Also,

the existing methods concerned about the misalignments

between iris patches (e.g., [34]) approached the problem

from parametric perspectives, and working in the polar

representations of the iris data.

1.1. Contributions

This paper we describe one iris recognition algorithm

that offers three contributions:

• we describe a novel segmentation-less polar represen-

tation of the iris data. Even though this representation

only partially attenuates the differences in pupillary di-

lation and scale, it provides novel types of informa-

tion that are useful for biometric recognition (e.g., the

shape of the pupil);

• we propose a processing chain that depends uniquely

of a bounding box ([x, r] ∈ N3) provided by an iris

detector (as a preprocessing step), which is in oppo-

sition to previously published non-holistic recognition

algorithms;

• the proposed algorithm is agnostic to the levels of

pupillary dilation, which is also a novelty;

Finally, the use of deep-learning frameworks to infer bio-

logically corresponding patches should also be considered a

contribution. Note that the concept of corresponding patch

lies in a hyperspace of much lower dimensionality than the

space of identities, where deep learning frameworks typ-

ically operate. As such, the amount of labelled data re-

quired for appropriate learning of our model is kept rela-

tively short.

1.2. Iris Recognition: Strides and Challenges

Strides in iris recognition have been concentrated in: 1)

extending the data acquisition volume; 2) improving robust-

ness to unconstrained conditions; 3) obtaining interpretable

iris representations; 4) developing cancellable signatures;

and 5) providing cross-sensor operability. In terms of the

data acquisition volume, the iris-on-the-move system [19]

can be highlighted, along with Hsieh et al.’s [11] work,

based in wavefront coding and super-resolution techniques.

Regarding the unconstrained recognition problem, Dong et

al. [7] used adaptive matching to identify the best features

per individual. Pillai et al. [23] sparsely represented ran-

domly projected iris patches, while Yang et al. [36] relied

in high-order information to match irises pairs, and Alonzo-

Fernandez et al. [8] proposed a super-resolution method

based on eigen-transformations of the iris patches. Bit con-

sistency is a concern, with approaches selecting only parts

of the biometric signatures for matching (e.g. [12], [33]

and [17]). Recognition in mobile devices is a particular

case of the unconstrained recognition problem, with two

international evaluation contests being recently conducted:

using visible wavelength (Marsico et al. [18]) and NIR data

(Zhang et al. [40]). In terms of interpretability, Chen et

al. [3] exclusively relied in features that are intuitively un-

derstood by humans. For cross-sensor recognition, Pillai et

al. [24] learned kernel-based transformations between the

data acquired by different devices. Finally, Zhao et al. [37]

suggested the concept of negative recognition as a way to

cancellable biometrics, using complementary information

of the biometric data for matching.

Deep learning-based methods have been advancing

the state-of-the-art, with even fine-tuning providing good

results [21]. Solutions have been proposed either for

particular phases of the recognition chain (e.g., segmen-

tation [13] or spoofing detection [20]), or for the whole

process: Gangwar and Joshi [9] claimed that very deep

architectures are particularly efficient to model the micro-

structures of iris texture. Zhang et al. [41] developed

a feature fusion strategy that considers the complemen-

tarity of the iris and periocular information. Zhao and

Kumar [38] used fully convolutional networks to generate

phase-based discriminative features, based on an extended

triple loss that accounts for bit shifting and non-iris regions

masking. Finally, Proenca and Neves [25] analyzed the

phase/magnitude of the free-form deformation fields (2D

registration vectors) between images, using second-order

statistics to discriminate between genuine and impostor

pairs.

The remainder of this paper is organized as follows:

Section 2 provides a detailed description of the proposed

method. In Section 3 we discuss the obtained results and

the conclusions are given in Section 4.

Segmentation [30]

Occlusions Detect.

Learning Sample Segmented Noise-Free Data

Normalized Data

Normalized Correspondences

Cartesian Correspondences

Polar Correspondences

Normalization [5]

Manual Annotation

fn→c(x) fc→p(x)

x1,x2

.

.

.

[ ]

n

x1,x2

.

.

.

[ ]

c

x1,x2

.

.

.

[ ]

p

Registration CNN

(VGG-19, Learning)

+

+

+

+

+

+

+

+

+

+

+

+

Gallery + Probe

SSD [15]

Detected Irises

fc→p(x)

Registration CNN (VGG-19)

Classification CNN(VGG-19, Learning)

. . .. . .

Registration MapsClassification CNN

(VGG-19)

✓✗

A

B

C

Test

Figure 2. Cohesive perspective of the recognition method proposed in this paper: we accurately segment and normalize the noise-free iris

regions of a learning set, and manually annotate the regions in genuine pairs that appear to regard the same biological region corresponding

iris patches. Such correspondences are remapped to Cartesian coordinates and - next - to a pseudo-polar coordinate system that doesn’t

require image segmentation, where a CNN learns the notion of corresponding region.

2. Proposed Method

2.1. Learning Phase

We employ the term biologically corresponding iris

patches to refer regions in pairs of genuine images that, un-

der visual inspection, seem to regard the same biological

region.

2.1.1 Biologically Corresponding Iris Patches

Let I = {I1, . . . , In} be a learning set of n iris images. For

every I. we obtain its shape-flexible and size-invariant nor-

malized representation, as proposed by Daugman [5]. The

mapping is given by I(

x(r, θ), y(r, θ))

→ I(r, θ):

[

x(r, θ)y(r, θ)

]

=

[

xp(θ) xs(θ)yp(θ) xs(θ)

] [

1− r

r

]

, (1)

with the dimensionless parameter r spanning in the unit in-

terval.

Next, for pairs of genuine normalized images {I, I ′}, we

manually annotate the control points that seem to regard the

same iris feature. Groups of control points define two con-

vex polygons Γ and Γ′, represented by the coloured dots

(xi and x′i) in Fig. 3. Let xi = (xi, yi) and x′

i = (x′i, y

′i),

i = {1, . . . t} be the locations of such pointwise correspon-

dences. As in [25], we learn two functions f1, f2 that es-

tablish a dense set of correspondences between positions

(rows/columns) in Γ and Γ′, f1, f2 : N2 → N, such that

∀ x′i ∈ Γ′, x′

i =(

fc(x), fr(x))

:

f2(x) = λTc [φ, p(x)], (2)

f1(x) = λTr [φ, p(x)], (3)

with φ =[

φ(

|x−x1|2)

, . . . , φ(

|x−xt|2)]

, |.|2 represent-

ing the ℓ2 norm, φ(r) = e(−r/κ)2 being a radial basis func-

tion and p(x) = [1, x, y] being a polynomial basis of first

degree for a 2-dimensional vector space (κ = 0.1 was used

in all our experiments). To obtain the λ coefficients, we de-

fine a t×t matrix A, Ai,j = φ(

|xi−xj |2)

and P as the t×3polynomial basis matrix, such that P = [p(x1); . . . ; p(xt)].Then, λc and λr are given by:

λc =

[

A P

P T 0

]−1 [x′c

0

]

(4)

λr =

[

A P

P T 0

]−1 [x′r

0

]

(5)

with x′c=[x′

1, . . . , x′t]T and x′

r=[y′1, . . . , y′t]T concatenating

the horizontal (column) and vertical (row) positions of the

control points in Γ′.

According to this procedure, we deem that any position

x ∈ Γ corresponds biologically to x′ =(

fc(x), fr(x))

∈Γ′. As Γ and Γ′ have different size and shape, this set

of correspondences implicitly encodes the non-linear defor-

mations that affect the iris texture, which is the core for the

performance of the proposed method.

x1

x′

1

x2x

′

2

x3

x′

3

x4

x′

4

Γ Γ′

x

α β

γ

δ

x′

α′

β′

γ′

δ′

αγ ≈ α′

γ′, βδ ≈ β′

δ′

Figure 3. Biologically corresponding iris patches. Based on a set

of manually annotated corresponding control points (x. and x′

.) in

a pair of iris samples, two polygons (Γ and Γ′) are defined. For

every position inside Γ, the corresponding position in Γ′ is found,

with both points deemed to refer to the same biological region.

2.1.2 Segmentation-less Polar Representation

Next, we convert all positions in the normalized iris data

into a segmentation-less polar space. Let x = (x, y) be

a point in the h × w normalized representation of the iris.

As an intermediate step, we map x into the original Carte-

sian space and from there into the segmentation-less space.

Function fn→c(.) : N2 → N

2, performs the first mapping:

fn→c(x) = ρ xp(θ) + (1− ρ) xs(θ), (6)

with xp and xs denoting the coordinates of the pupillary

and scleric borders at angle θ, and (θ, ρ) given by:

θ =2π(x− 1)

w, ρ =

h− y

h− 1. (7)

Next, using fc→p(x) : N2 → N2, we map the position

in the Cartesian space into the polar representation:

fc→p(x) =[

atan2( s

2 − y

x− s2

)

w(2π)−1 + 1,

(|x− [s

2,s

2]|2 − τp)h(xc − τp)

−1]

, (8)

with rp denoting an average radius for the pupils and s× s

is the size of the bounding box returned by an iris detection

step. In this sense, fc→p(fn→c(x)) converts points in the

normalized iris data into the segmentation-less polar repre-

sentation.

Finally, we consider patches P (from Γ) and P ′ (from

Γ′) of 15×15 pixels, centered at each point correspondence.

This set of patches is concatenated into a set of negative in-

stances (non-corresponding patches), feeding the learning

process of a Convolutional Neural Network (”Registration

CNN” ), responsible to discriminate between patches that

correspond or not (part ”B” in Fig. 2). The architecture of

this CNN is based in the VGG-19 model, adapting the size

of the input layer to depth 2, and removing all the pooling

layers, due to the relatively small dimensions of the input

data. The output layer is a soft max loss corresponding to

the probability of two iris patches correspond. Learning

was done according to the stochastic gradient descend algo-

rithm, with an initial learning rate of 1e−2, momentum set

to 0.9 and weight decay equals to 1e−3 .

The steps of the learning phase are shown in Fig. 2 as

the ”C” group. We use the popular SSD [13] detector in the

learning data, to detect the irises in both images. The detec-

tor was adapted so to use default boxes with a single ratio,

corresponding to squares. This way, the detector returns a

parameterization [xc, s] ∈ N3, with xc being the center of

the s× s detected region.

Using fc→p(x), ∀x we get the segmentation-less polar

representation of the iris, where a τr × τc grid of equally

spaced points is overlapped. Let I, I′ be a pair of irises in

the segmentation-less polar space. Let x be the position of

one grid point in the first image. We define a 15× 15 patch

around this point and densely sample the x ± [δr, δc] po-

sitions (δr ∈ {1, . . . ,∆r}, δc ∈ {1, . . . ,∆c}) in the other

image, and take 15 × 15 patches centered at each position.

Pairs of patches are provided to the previously learned Reg-

istration CNN, with output scores directly corresponding to

the probability that pairs of patches in the first and second

image correspond biologically. By rearranging the CNN

scores, we create the so-called registration maps that pro-

vide information about the registration scores for a specific

direction and magnitude of deviation. The registration maps

from the learning set are used as input of the ”Classification

CNN”, also based in VGG-19 architecture, that is respon-

sible to discriminate between genuine and impostor maps.

Once again, the output layer is a soft max loss corresponding

to the probability of a set of registration maps regarding a

genuine comparison. As previously, learning was done ac-

cording to the stochastic gradient descend algorithm, with

an initial learning rate of 1e−2, momentum set to 0.9 and

weight decay equals to 1e−3 .

2.2. Classification Phase

The classification (recognition) processing chain is en-

closed by the red dashed rectangle in Fig. 2. For a

query+gallery pair of iris images, the detection and map-

ping process is carried out in an exactly equal way to the

used in the ”C” group of the learning phase. We use the

SSD [13] detector to get information about the iris center

and size [xc, s] and obtain the segmentation-less polar rep-

resentation of the iris. Finally, according to the overlapped

grid control points, we obtain the registration maps that feed

the learned classification CNN model, which returns the

probability (output of a softmax layer) that the query and

the gallery regard the same identity.

3. Results and Discussion

3.1. Experimental Setting and Preprocessing

Experiments were conducted in three datasets: CASIA-

IrisV4-Lamp, CASIA-IrisV4-Thousand 1 and WVU 2, with

examples of the predominant degradation factors given in

Fig. 4: off-angle and partially occluded irises, glasses, ex-

tremely dilated/constricted pupils (in all sets) and local illu-

mination variations (in WVU). 822 classes (411 eyes) from

the CASIA-IrisV4-Lamp set, 2,000 classes (1,000 eyes)

from the CASIA-IrisV4-Thousand and 638 classes from the

WVU set were considered. For the CASIA sets, 10 images

per class were used, while for the WVU set the number of

images per class varied between 2 and 10. Identities were

separated into two halves for learning/test (the first half of

the identities, according to the alphanumeric order of the

filenames, was chosen for learning), with the resulting data

enabling to perform 18,495 (genuine) + 8,425,500 (impos-

tor) pairwise comparisons for the CASIA-IrisV4-Lamp set,

45,000 (gen.) + 49,950,000 (imp.) for the CASIA-IrisV4-

Thousand and 11,301 (gen.) + 7,470,767 (imp.) for the

WVU set. In every experiment, the number of impostor

pairwise comparisons considered was limited to 1,000,000

(randomly chosen).

Using a coarse-to-fine segmentation strategy [30], com-

posed of a form fitting step and a geodesic active contours

algorithm, images were segmented, with the iris boundaries

being described by shapes of 20 degrees-of-freedom (dof)

and the scleric boundary described by 3 dof. The accuracy

of the segmentation step was manually verified, and differ-

ent parameterizations considered, to assure that all the sam-

ples were appropriately segmented. The ambitious open-

world setting was the unique considered, i.e., all images

from half of the identities in each set were used for learn-

ing purposes, with the recognition experiments being con-

1CASIA iris image database, http://biometrics.

idealtest.org2West Virginia University iris dataset, http://www.clarkson.

edu/citer/research/collections/

ducted in previously unseen identities.

Figure 4. Datasets used in our experiments: from top to bot-

tom rows, images of the CASIA-IrisV4-Lamp, CASIA-IrisV4-

Thousand and WVU sets are shown.

As baselines, five methods were considered to represent

the state-of-the-art : 1) IRINA [25], which analyses in a

non-holistic way the correspondences between iris patches

in the segmented and normalized irises; 2) the work due

to Yang et al. [36] (using the O2PT iris-only variant, with

block size w = 2, h = 14, translation vector [6, 3]T and

neighbourhood 8× 8); 3) Sun and Tan’s [31] method (with

di-lobe and tri-lobe filters, Gaussians 5× 5, σ = 1.7, inter-

lobe distances {5,9} and sequential feature selection); 4) the

keypoints-based method due to Belcher and Du [1] (with 64

bins = 4 (horizontal) × 4 (vertical) × 4 (orientation), SIFT

descriptors extracted using VLFeat package 3 ); and 5) the

OSIRISv4 [22] framework, to represent the classical pro-

cessing chain proposed by Daugman [5] (using the Viterbi

algorithm for segmentation, Gabor filters for feature extrac-

tion and the Hamming distance for matching).

Three performance measures are reported: the decid-

ability index (d′), the area under curve (AUC) and the

equal error rate (EER), also providing the corresponding re-

ceiver operating characteristic (ROC) curves. The pairwise

comparisons per dataset were divided into random samples

(drew with repetition), each one with 90% of the available

pairs. Then, tests were conducted iteratively in each sam-

ple, with the results considered to approximate the confi-

dence intervals at each point, in a bootstrapping-like strat-

egy. To facilitate the reproducibility of the experiments re-

ported here, we release the MatConvNet [35] trained ”Reg-

istration” and ”Classification” VGG-19 models 4.

3.2. Parameter Tuning

The performance of our method depends of three param-

eters: 1) the minimum probability τ1 required for relative

positions in the eye patches being ”iris” to be considered in

the analysis (p(”Iris” | x) > τ1, x = [x, y]); 2) the size of

the grid (τr × τc) overlapped in the polar representations of

the data; and 3) the neighbourhood ±(∆r,∆c) considered

when searching for biologically corresponding patches.

3http://www.vlfeat.org/4http://tinyurl.com/y3e6l73r

1

0.5

0

a) p(”Iris” | x)

0 0.5 1

0

0.1

0.2

0.3

b) Radii pupil/scleraP

rob.

-10 -5 0 5

0

0.2

0.4

c) Normal. ∆row

Pro

b.

-10 0 10

0

0.2

0.4

d) Normal. ∆col

Pro

b.

-20 0 20

0

0.2

0.4

0.6

e) Polar ∆row

Pro

b.

-10 0 10

0

0.2

0.4

0.6

f) Polar ∆col

Pro

b.

Figure 5. a) Probability that a pixel inside the detected ”Eye” re-

gion is noise-free and iris; b) Relative radii between the pupil

and the sclera; c)-d) rows/columns offset between corresponding

points in the normalized images (in pixels); e)-f) corresponding

offsets in the segmentation-less polar images.

In Fig. 5 we provide the p(”Iris” | x) values for the

learning data. For each bounding box parameterized by

[x, r] ∈ N3 (i.e., x → x + [r, r]) we used the segmen-

tation information to classify each position into ”iris” or

”non-iris”. It can be seen that there is a residual probability

of observing iris pixels inside a central circle (correspond-

ing to the pupil) and at the image corners (corresponding

to sclera). Also, the upper part of the iris is often occluded

by eyelids/eyelashes, which reduces the ”iris” probability

in these regions. Setting τ1 = 0, a binary mask B can be

obtained:

B(x) =

{

1, if p(”Iris”| x) > τ10, otherwise

Using the circular Hough transform, two concentric cir-

cumferences (xp, rp) and (xs, rs) were fitted to B, and

deemed to be the generic pupillary and scleric bound-

aries. These parameterizations are the key to obtain the

segmentation-less polar representations of the iris, both in

the learning and test sets. For any image in these sets, we

assume that (xp, rp) and (xs, rs) correspond to the pupil-

lary and scleric boundaries and normalize the data, accord-

ing to (1).

In terms of the (τr × τc) values, they determine the

amount of information extracted from the iris, and in-

crease quadratically the time complexity of the encoding

step. Keeping τrτc

= hw = 1

2 , we assessed the per-

formance of our method in a CASIA-4-Lamp validation

set with τr = {2, 4, 8, 16, 32}, obtaining AUC values

{0.980, 0.986, 0.995, 0.995, 0.995} that support the choice

τr = 8. Finally, the (∆r,∆c) values were tuned according

to the information taken from the learning data provided in

Fig. 5 b)-f). It can be seen that, comparing the relative mis-

alignments between the biologically corresponding points

in the normalized (c) and d)) and in the polar (e) and f))

representations of the iris data, the major differences regard

the ∆r values, i.e., polar representations tend to provide

larger row misalignments than the normalized representa-

tions, justifying the (∆r = 10,∆c = 5) choice.

3.3. Patch­Based vs. Holistic CNNs

At a first level, we compared the performance of the

method proposed in this paper to the classical (holistic)

ways deep learning classification strategies have been used

in biometrics research (CNNs receiving pairs of samples

and discriminating between the genuine and impostor com-

parisons). Four CNNs were considered: 1) receiving the

whole ocular samples represented in the Cartesian space

(denoted as CNN-Holistic; 2) receiving the cropped iris re-

gions, in the same Cartesian space (CNN-Iris); 3) using

the segmented and normalized representation of iris data

(CNN-Normalized); and; 4) using the segmentation-less po-

lar representations proposed in this paper (CNN-Polar). As

data augmentation techniques, two label-preserving trans-

formations were used: 1) to simulate the scale and trans-

lation samples inconsistency, patches of scale [0.85, 0.95](values drew uniformly) were randomly cropped from the

learning set; and 2) as an intensity transform, we obtained

the principal components of the pixels intensities in the

learning data and synthesized images by adding to each

pixel multiples of the largest eigenvectors, with magnitude

equal to the corresponding eigenvalues [14]:

x(new) = x(old) + [v1,v2,v3](

α⊙ [λ1, λ2, λ3]T)

, (9)

with ⊙ denoting the element-wise multiplication, v and λ

denoting the eigenvectors and eigenvalues of the learning

data covariance matrix and α ∈ R3 being randomly drew

from the Gaussian N (0, 0.1).

The results are provided in Fig. 6 and turn clear the

advantages of the non-holistic analysis when compared

to its competitors. The improvements in performance

with respect to the runner up strategy (invariably CNN-

Normalized) were observed for all datasets, increasing the

AUC values between 5% to 15%. Among the holistic CNN

frameworks, using the segmented and normalized represen-

tations of the iris provided the best performance. It is in-

teresting to note that the segmentation-less polar represen-

tation got the runner-up performance in the CASIA-IrisV4-

Lamp and CASIA-IrisV4-Thousand datasets and even at-

tained the best results among all holistic representations in

the WVU set. In all cases, these normalized representations

provide better performance than when using the cropped

CASIA-IrisV4-Lamp

FP

TP

CASIA-IrisV4-Thousand

FP

TP

WVU

FP

TP

Proposed CNN-Holistic CNN-Iris CNN-Normal. CNN-Polar

Figure 6. ROC curves of the proposed method and of four holis-

tic CNNs typically used in biometrics: using the ocular image

(CNN-Holistic); the cropped iris (CNN-Iris); the normalized iris

data (CNN-Normalized); or the segmentation-less polar represen-

tation (CNN-Polar) as input.

eye data in the Cartesian domain and - even more evidently

- than using the whole ocular region as input of the CNN.

Method AUC d’ EER

CASIA-IrisV4-Lamp

Proposed Method 0.999 ± 5e−4 6.773 ± 0.482 0.006 ± 0.002

IRINA [25] 0.983 ± 8e−4 3.805 ± 0.457 0.038 ± 0.005

Sun and Tan [31] 0.992 ± 4e−4 4.448 ± 0.404 0.029 ± 0.005

Yang et al. [36] 0.993 ± 5e−4 4.629 ± 0.385 0.028 ± 0.004

OSIRIS [22] 0.992 ± 4e−4 4.017 ± 0.490 0.031 ± 0.006

Belcher and Du [1] 0.938 ± 0.007 2.933 ± 0.696 0.097 ± 0.011

CASIA-IrisV4-Thousand

Proposed Method 0.981 ± 7e−4 6.611 ± 0.484 0.030 ± 0.011

IRINA [25] 0.961 ± 8e−4 3.600 ± 0.488 0.051 ± 0.012

Sun and Tan [31] 0.964 ± 1e−3 4.095 ± 0.583 0.042 ± 0.006

Yang et al. [36] 0.978 ± 7e−4 4.995 ± 0.366 0.035 ± 0.004

OSIRIS [22] 0.967 ± 0.001 3.471 ± 0.560 0.058 ± 0.007

Belcher and Du [1] 0.901 ± 0.006 2.104 ± 0.597 0.084 ± 0.011

WVU

Proposed Method 0.969 ± 0.001 4.850 ± 0.500 0.063 ± 0.007

IRINA [25] 0.931 ± 0.002 4.066 ± 0.489 0.081 ± 0.007

Sun and Tan [31] 0.917 ± 0.002 2.552 ± 0.193 0.098 ± 0.007

Yang et al. [36] 0.970 ± 0.002 5.210 ± 0.185 0.055 ± 0.008

OSIRIS [22] 0.903 ± 0.005 2.801 ± 0.382 0.073 ± 0.009

Belcher and Du [1] 0.882 ± 0.011 2.008 ± 0.780 0.166 ± 0.015

Table 1. Performance summary of the proposed method with re-

spect to three state-of-the-art methods.

3.4. State­of­the­Art Comparison

Performance comparison with respect to the baselines

is shown in Fig. 7. Overall, the proposed method got ei-

ther the best or the runner-up performance. Results were

particularly good for the CASIA-IrisV4-Lamp, due to hav-

ing learned the ”Classification CNN” in (learning) pairwise

samples of that set. In all cases, the approach due to Belcher

and Du got the worst results, reinforcing the conclusion that

keypoint-based approaches fail in iris recognition, due to

the non-linear deformations that occur in the iris texture as a

result of pupillary dilation. In most cases, Yang et al.’s per-

formance was slightly better than ours’ in the high-FAR re-

gion of the performance space. When compared to IRINA,

it is evident that the method proposed in this paper repre-

sents an improvement in performance, essentially due to the

use of deep-learning classification techniques to discrimi-

nate between the genuine and impostors registration maps.

The open-world setting is a challenge for IRINA’s match-

ing phase, considering that the observed performance of this

method was far below the reported in the original publica-

tion. Finally, Sun and Tan and OSIRIS’ performance values

were close to the optimal value in the highest-FAR regions

of the performance space, but invariably diverged of ours

and Yang et al.’s results in the lowest-FAR regions.

Finally, as an insight for potential ways to improve per-

formance, in Fig. 8 we provide examples of the worst gen-

CASIA-IrisV4-Lamp

FP

TP

CASIA-IrisV4-Thousand

FP

TP

WVU

FP

TP

Proposed IRINA [25] Sun and Tan [31]

Yang et al. [36] OSIRIS [22] Belcher and Du [1]

Figure 7. Comparison between the ROC curves observed for the

proposed method, when compared to five techniques considered

to represent the state-of-the-art.

uine (framed in green) and impostor (framed in red) com-

parisons, listing the factors that - under visual perception-

were deemed to justify such poorest scores. Overall, the

proposed method yields poor genuine scores when the bot-

tom part of the iris is severely occluded by eyelids, whereas

the worst impostor scores yielded from extremely dilated

pupils that produced segmentation-less polar representa-

tions largely occupied by the pupillary regions.

l l l l l l

✗

Non-circular pupils

Heavy occlusionsLack iris minutia ✗

Extremely dilated pupils

Heavy occlusions

Off-angle gaze

Figure 8. Example of mage pairs that produced the worst genuine

(framed in green) and impostor (framed in red) matching scores in

our method.

4. Conclusions and Further Work

Considering the performance of deep learning-based so-

lutions in biometrics and the requirements of this kind

of frameworks in terms of labelled data for appropriately

learning identities, in this paper we proposed an iris recog-

nition algorithm that uses the power of deep learning ar-

chitectures exclusively to infer the misalignments between

biologically corresponding regions in pairs of iris images

represented in a segmentation-less polar domain. Based on

the homogeneity and magnitude of the deviations, it is rela-

tively easy to distinguish between the genuine and impostor

comparisons, yielding a recognition strategy that achieves

state-of-the-art performance, at a much less demanding vol-

ume of learning data than its competitors, without even re-

quiring data to be segmented and agnostic to the levels of

pupillary dilation. Our experiments turn evident that this

algorithm is particularly valuable in case of low quality

data, where accurate segmentation becomes a too challeng-

ing task. As main weakness, the computational complexity

of the matching process should be pointed out, as the devi-

ation maps have to be explicitly obtained for each pairwise

comparison. We are now concentrated in finding alternate

strategies to obtain the 2D deformation maps and reduce the

computational cost of matching.

Acknowledgements

This work is funded by FCT/MEC through national

funds and co-funded by FEDER - PT2020 partnership

agreement under the projects UID/EEA/50008/2019 and

POCI-01-0247-FEDER-033395.

References

[1] C. Belcher and Y. Du. Region-based SIFT approach to iris

recognition. Opt Lasers Engeneering, vol. 47, no. 1, pag.

139–147, 2009. 5, 7, 8

[2] K. Bowyer, K. Hollingsworth and P. Flynn. A Survey of

Iris Biometrics Research: 2008-2010. In M. J. Burge and K.

W. Bowyer (eds.), Handbook of Iris Recognition, Advances

in Computer Vision and Pattern Recognition, pag. 15–54,

Springer, 2013. 1

[3] J. Chen, F. Shen, D. Chen and P. Flynn. Iris Recognition

Based on Human-Interpretable Features. IEEE Transactions

on Information Forensics and Security, vol. 11, no. 7, pag.

1476–1485, 2016. 2

[4] J. Daugman. High Confidence Visual Recognition of Persons

by a Test of Statistical Independence. IEEE Transactions on

Pattern Analysis and Machine Intelligence, vol. 15, no. 11,

pag. 1148–1161, 1993. 1

[5] J. Daugman. New Methods in Iris Recognition. IEEE Trans-

actions on Systems, Man, and Cybernetics - part B: Cyber-

netics, vol. 37, no. 5, pag. 1167–1175, 2004. 3, 5

[6] C. Ding and D. Tao. Robust Face Recognition via Mul-

timodal Deep Face Representation. IEEE Transactions on

Multimedia, vol. 17, no. 11, pag. 2049–2058, 2015. 1

[7] W. Dong, Z. Sun and T. Tan. Iris matching based on person-

alized weight map. IEEE Transactions on Pattern Analysis

and Machine Intelligence, vol. 33, no. 9, pag. 1744–1757,

2011. 2

[8] F. Alonso-Fernandez, R. Farrugia and J. Bigun. Eigen-Patch

Iris Super-Resolution for Iris Recognition Improvement. In

proceedings of the 23rd European Signal Processing Confer-

ence, doi: 10.1109/EUSIPCO.2015.7362348, 2015.

2

[9] A. Gangwar and A. Joshi. DeepIrisNet: Deep Iris Repre-

sentation With Applications in Iris Recognition and Cross-

Sensor Iris Recognition. In proceedings of the International

Conference on Image Processing, pag. 2301–2305, 2016. 2

[10] S. Gao, Y. Zhang, K. Jia, J. Lu and Y. Zhang. Single Sam-

ple Face Recognition via Learning Deep Supervised Autoen-

coders. IEEE Transactions on Information Forensics and Se-

curity, vol. 10, no. 10, pag. 2108–2118, 2015. 1

[11] S-H. Hsieh, Y-H. Li, C-H. Tien and C-C. Chang. Extend-

ing the Capture Volume of an Iris Recognition System Using

Wavefront Coding and Super-Resolution. IEEE Transactions

on Cybernetics, doi: 10.1109/TCYB.2015.2504388,

2016. 2

[12] Y. Hu, K. Sirlantzis and G. Howells. Exploiting stable and

discriminative iris weight map for iris recognition under less

constrained environment. In proceedings of the IEEE Inter-

national Conference on Biometrics Theory, Applications and

Systems, pag. 1–8, 2015. 2

[13] N. Liu, H. Li, M. Zhang, J. Liu, Z. Sun and T. Tan. Ac-

curate Iris Segmentation in Non-cooperative Environments

Using Fully Convolutional Networks. In proceedings of the

International Conference on Biometrics, pag. 1–8, 2016. 2,

4, 5

[14] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet clas-

sification with deep convolutional neural networks. In pro-

ceedings of the Advances in Neural Information Processing

Systems Conference, pag. 1097–1105, 2012. 1, 6

[15] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C-

Y. Fu and A. Berg. SSD: Single Shot Multibox Detector.

arXiv:1512.02325, 2016. 3

[16] J. Long, E. Schelhamar and T. Darrell. Fully Convolutional

Networks for Semantic Segmentation. In proceedings of the

IEEE Conference on Computer Vision and Pattern Recogni-

tion, pag. 640–651, 2015. 1

[17] N. Mahadeo, A. Paplinski and S. Ray. Optimization of Iris

Codes for Improved Recognition. In proceedings of the

IEEE Conference on Computer Vision and Pattern Recog-

nition Workshops, pag. 48–55, 2014. 2

[18] M. De Marsico, M. Nappi and H. Proenca Results from

MICHE II: Mobile iris challenge evaluation II. Pattern

Recognition Letters, vol. 91, no. 1, pag. 3–10, 2017. 2

[19] J. Matey, O. Naroditsky, K. Hanna, R. Kolczynski,

D. Lolanoco, S. Mangru, M. Tinker, T. Zappia and W. Zhao.

Iris on the move: Acquisition of images for iris recognition

in less constrained environments. Proceedings of the IEEE,

vol. 94, no. 11, pag. 1936–1947, 2006. 2

[20] D. Menotti, G. Chiachia, W. Schwartz, H. Pedrini, A. Falcao

and A. Rocha. Deep Representations for Iris, Face, and Fin-

gerprint Spoofing Detection. IEEE Transactions on Infor-

mation Forensics and Security, vol. 10, no. 4, pag. 864–879,

2015. 2

[21] K. Nguyen, C. Fookes, A. Ross and S. Sridharan. Iris recog-

nition with off-the-shelf CNN Features: A deep learning per-

spective. IEEE Access, vol. 6, pag. 18848–18855, 2017. 2

[22] N. Othman, B. Dorizzi and S. Garcia-Salicetti. OSIRIS: An

open source iris recognition software. Pattern recognition

Letters, vol. 82, no. 2, pag. 124-131, 2016. 5, 7, 8

[23] J. Pillai, V. Patel, R. Chellappa and N. Ratha. Secure and

robust iris recognition using random projections and sparse

representations. IEEE Transactions on Pattern Analysis and

Machine Intelligence, vol. 33, no. 9, pag. 1877–1893, 2011.

2

[24] J. Pillai, M. Puertas and R. Chellappa. Cross-sensor iris

recognition through kernel learning. IEEE Transactions on

Pattern Analysis and Machine Intelligence, vol. 36, no. 1,

pag. 73–85, 2014. 2

[25] H. Proenca and J. Neves. IRINA: Iris Recognition (even) in

Inaccurately Segmented Data. In proceedings of the Interna-

tional Conference on Computer Vision and Pattern Recogni-

tion, pag. 6747–6756, 2017. 2, 3, 5, 7, 8

[26] H. Proenca and J. Neves. Deep-PRWIS: Periocular Recogni-

tion Without the Iris and Sclera Using Deep Learning Frame-

works. IEEE Transactions on Information Forensics and Se-

curity, vol. 13, no. 4, pag. 888–896, 2018. 1

[27] H. Qin and M. El-Yacoubi. Deep Representation-Based Fea-

ture Extraction and Recovering for Finger-Vein Verification.

IEEE Transactions on Information Forensics and Security,

vol. 12, no. 8, pag. 1816–1829, 2017. 1

[28] A. Rodriguez, J. Panza and B. Vijaya Kumar, Segmentation-

free Ocular Detection and Recognition. In proceedings of

the SPIE Conf. on Biometrics for Human Identification, vol.

8029B, 2011. 2

[29] A. Rodriguez and B. Vijaya Kumar, Segmentation-free bio-

metric recognition using correlation filters. Academic Press

Library in Signal Processing: Image, Video Processing and

Analysis, Hardware, Audio, Acoustic and Speech Processing,

vol. 4, pag. 403–460 2013. 2

[30] S. Shah and A. Ross. Iris Segmentation Using Geodesic Ac-

tive Contours. IEEE Transactions on Information Forensics

and Security, vol. 4, no. 4, pag. 824–836, 2009. 3, 5

[31] Z. Sun and T. Tan. Ordinal Measures for Iris Recognition.

IEEE Transactions on Pattern Analysis and Machine Intelli-

gence, vol. 31, no. 12, pag. 2211–2226, 2009. 5, 7, 8

[32] C. Szegedy, A. Toshev and D. Erhan. Deep Neural Networks

for Object Detection. In proceedings of the Advances in Neu-

ral Information Processing Systems Conference, pag. 2553–

2561, 2013. 1

[33] C-W. Tan and A. Kumar. Accurate Iris Recognition at a Dis-

tance Using Stabilized Iris Encoding and Zernike Moments

Phase Features. IEEE Transactions on Image Processing,

vol. 23, no. 9, pag. 3962–3974, 2014. 2

[34] J. Thornton, M. Savvides and B.V.K. Vijaya Kumar,. A uni-

fied Bayesian approach to deformed pattern matching of iris

images. IEEE Transactions on Pattern Analysis and Machine

Intelligence, vol. 29, no. 4, pag. 596–606, 2007. 2

[35] A. Vedaldi and K. Lenc. MatConvNet: Convolutional Neural

Networks for MATLAB. In proceedings of the 23rd ACM in-

ternational conference on Multimedia, pag. 689–692, 2015.

5

[36] G. Yang, H. Zeng, P. Li and L. Zhang. High-Order Infor-

mation for Robust Iris Recognition Under Less Controlled

Conditions. In proceedings of the International Conference

on Image Processing, pag. 4535–4539, 2015. 2, 5, 7, 8

[37] D. Zhao, W. Luo, R. Liu and L. Yue. Negative Iris Recog-

nition. IEEE Transactions on Dependable and Secure Com-

puting, doi: 10.1109/TDSC.2015.2507133, 2016. 2

[38] Z. Zhao and A. Kumar. Towards More Accurate Iris Recog-

nition Using Deeply Learned Spatially Corresponding Fea-

tures. In proceedings of the International Conference on

Computer Vision, pag. 3829–3838, 2017. 2

[39] Z. Zhao and A. Kumar. Improving Periocular Recognition by

Explicit Attention to Critical Regions in Deep Neural Net-

work. IEEE Transactions on Information Forensics and Se-

curity, vol. 13, no. 12, pag. 2937–2952, 2018. 1

[40] M. Zhang, Q. Zhang, Z. Sun, S. Zhou and N. Ahmed. The

BTAS competition on mobile iris recognition. In proceed-

ings of the IEEE International Conference on Biometrics

Theory, Applications and Systems, pag. 1–7, 2016. 2

[41] Q. Zhang, H. Li, Z. Sun and T. Tan. Deep Feature Fusion

for Iris and Periocular Biometrics on Mobile Devices. IEEE

Transactions on Information Forensics and Security, vol. 13,

no. 11, pag. 2897–2912, 2018. 2